[Intel-gfx] [RFC 01/12] drm/i915: Expose list of clients in sysfs

[Tvrtko Ursulin]

On 10/03/2020 00:13, Chris Wilson wrote:
> Quoting Tvrtko Ursulin (2020-03-09 23:26:34)
>>
>> On 09/03/2020 21:34, Chris Wilson wrote:
>>> Quoting Tvrtko Ursulin (2020-03-09 18:31:18)
>>>> +struct i915_drm_client *
>>>> +i915_drm_client_add(struct i915_drm_clients *clients, struct task_struct *task)
>>>> +{
>>>> +       struct i915_drm_client *client;
>>>> +       int ret;
>>>> +
>>>> +       client = kzalloc(sizeof(*client), GFP_KERNEL);
>>>> +       if (!client)
>>>> +               return ERR_PTR(-ENOMEM);
>>>> +
>>>> +       kref_init(&client->kref);
>>>> +       client->clients = clients;
>>>> +
>>>> +       ret = mutex_lock_interruptible(&clients->lock);
>>>> +       if (ret)
>>>> +               goto err_id;
>>>> +       ret = xa_alloc_cyclic(&clients->xarray, &client->id, client,
>>>> +                             xa_limit_32b, &clients->next_id, GFP_KERNEL);
>>>
>>> So what's next_id used for that explains having the over-arching mutex?
>>
>> It's to give out client id's "cyclically" - before I apparently
>> misunderstood what xa_alloc_cyclic is supposed to do - I thought after
>> giving out id 1 it would give out 2 next, even if 1 was returned to the
>> pool in the meantime. But it doesn't, I need to track the start point
>> for the next search with "next".
> 
> Ok. A requirement of the API for the external counter.
>   
>> I want this to make intel_gpu_top's life easier, so it doesn't have to
>> deal with id recycling for all practical purposes.
> 
> Fair enough. I only worry about the radix nodes and sparse ids :)

I only found in docs that it should be efficient when the data is 
"densely clustered". And given that does appear based on a tree like 
structure I thought that means a few clusters of ids should be okay. But 
maybe in practice we would have more than a few clusters. I guess that 
could indeed be the case.. hm.. Maybe I could use a list and keep 
pointer to last entry. When u32 next wraps I reset to list head. 
Downside is any search for next free id potentially has to walk over one 
used up cluster. May be passable apart for IGT-type stress tests.

>> And a peek into xa implementation told me the internal lock is not
>> protecting "next.
> 
> See xa_alloc_cyclic(), seems to cover __xa_alloc_cycle (where *next is
> manipulated) under the xa_lock.

Ha, true, not sure how I went past top-level and forgot what's in there. :)

Regards,

Tvrtko